Operational fault diagnosis of autonomous underwater vehicles via a hybrid descriptor-temporal stacking framework

Reliable operation of Autonomous Underwater Vehicles (AUVs) in harsh marine environments depends on effective health monitoring and accurate awareness of their operating condition. However, AUV fault signatures are highly heterogeneous, involving both global statistical shifts and localized high-frequency temporal irregularities, which challenge traditional diagnostic models. To address these challenges and advance failure analysis in maritime systems, this paper proposes a novel hybrid diagnostic framework that combines physically interpretable handcrafted statistical descriptors with learned temporal representations. Using a leakage-free out-of-fold (OOF) stacking strategy, the framework integrates the global statistical characterization provided by a tree-based ensemble with the local temporal pattern learning of a residual 1D CNN. Evaluated on a publicly available experimental AUV benchmark, the proposed method achieves a test accuracy of 99.18% and a Macro-F1 score of 0.9914. To improve interpretability for practical maritime maintenance, feature importance and mean absolute SHAP analyses are applied to the descriptor-based branch. The results indicate that thruster-control variability and vehicle-attitude dynamics provide the largest aggregated physical contributions, supporting the plausibility of the learned diagnostic patterns. These findings demonstrate the potential of the proposed framework as an interpretable decision-support approach for AUV health monitoring and maintenance.